Today we’ve had the news that many have waited on for months.. Tesla’s Full Self Driving Beta Version 12 is now going out to customer cars. WholeMarsCatalog has shared on an X Space that he received a call from Tesla’s Autopilot team, asking him to confirm the VIN of his HW4 Tesla Model S.
During the call, the Autopilot team also warned that this beta release required additional caution, given a substantial architecture change in V12. They indicated the nag has been dialed up to ensure drivers ae paying extra attention.
The build is 2023.44.30.12 (FSD Beta V12.1.2) and removes around 300,000 lines of code, replaced by an end-to-end approach, which effectively takes video inputs in and controls out. This significant shift in Tesla’s approach to autonomous driving is potentially the unlock to overcome the long tail of edge cases found as cars navigate the dynamic and complex real-world environments.
The last time we seen V12 in action was back in August 2023 when Elon Musk live streamed the first drive on X. To release the build to the public at all, suggests the team are confident enough that V12 is close to the capabilities of V11 and may even be better in some circumstances.
During Musk’s drive, he showed new capability in V12, to go curb-to-curb, with the car parking itself at the conclusion of the route. WholeMarsCatalog is currently recording video of his first drive and this will be one of the important things to test. Many are also keen to understand if Actual Smart Summon has arrived with V12.
Here’s why V12 is important:
Paradigm Shift
- Traditional Approach: Previously, Tesla relied on a multi-stage pipeline with separate modules for perception, planning, and control. Each module processed sensor data and made decisions independently, adding complexity and potential for errors at handoffs.
- V12 Approach: V12 streamlines the process by using a single, massive neural network that takes raw camera data as input and directly outputs driving commands. This eliminates the need for hand-coded rules and intermediate representations, potentially leading to:
- More efficient and faster decision-making: The network can directly learn complex relationships between sensory input and driving actions, potentially enabling smoother and more natural maneuvers.
- Improved adaptability: The network can continuously learn and adapt to new scenarios based on real-world driving data, potentially leading to better performance in diverse environments.
Potential Benefits
- Enhanced Driving Performance: V12 has the potential to improve the smoothness, precision, and overall driving experience of Tesla’s Autopilot and FSD systems.
- Faster Development and Iteration: The end-to-end approach may allow Tesla to iterate on their self-driving software more quickly and efficiently, as changes made to the neural network can directly impact driving behavior.
- Reduced Hardware Dependency: While still requiring powerful hardware, V12’s efficiency could potentially enable self-driving capabilities on lower-cost hardware in the future.
Challenges and Concerns
- Explainability and Debugging: With a complex neural network, it can be challenging to understand why the system makes certain decisions, making debugging and error correction more difficult.
- Safety and Validation: Ensuring the safety of a system that learns directly from data is crucial. Extensive testing and validation are necessary to guarantee reliable and predictable behavior in all driving situations.
- Regulation and Public Perception: Public trust and regulatory approval are essential for the widespread adoption of self-driving cars. V12’s new approach may raise concerns that need to be addressed.
Update
Whole Mars Catalog has now posted the first videos from V12 on YouTube. Early impressions are good from the limited scenarios we see in the 2 videos, the car’s behaviour looks smooth, and human-like in its movements. A couple of notable changes are the speed at which it takes off from a stop, along with slowing for speed bumps to around 8MPH).
Sure, but not Australia right?